From a fundus image which contains affects of diabetic retinopathy, I have segmented part of each spot so that the segmented image contains pixels which identify the location of each spot in the original image. From the segmented image, I would like to return to the original image and capture the entire spot, not just part of it. It should be a problem with connected neighbors. Does anyone have experience in this area?
Attached is the original image and the segmented image.
@ Muhammad Zubair Ahmad The segmented image contains objects which represent only a portion of the objects in the original image. The goal is to return to the original image and capture all of the pixels within each object.
John, if you could clarify one thing for me. You are satisfied with the segmented image and now want to display those parts of the original image instead of the logical mask you have as a result of the segmentation algo
From what I understand, this is almost a supervised segmentation since that you spotted some area and want to grow the segmentation. Probably the easier way should be to apply a region growing. You can also use graph-cut which should perform well (http://www.eecs.berkeley.edu/~efros/courses/AP06/Papers/boykov-iccv-01.pdf). You have more recent references with the work of Boykov et al.
Maybe you will need to pre-process the image before to do these kind of segmentation to remove the vessels and highlight the interesting.
Check the work of L. Giancardo (https://scholar.google.com/citations?user=6x6xdCkAAAAJ&hl=en) which is probably the same area than what you are doing.
You could calculate the convex hull of all the points. If the convex hull is too big, you could allow for concavities with a maximum edge-length.
See Delaunay triangulation and Alpha-shapes.
Lambert Zijp, do you have some papers which applied this technique for classification?
John, sorry for a very late reply. I agree with Guillaume and further suggest that you look into the some techniques such as seeded region growing where you use the already detected points as seed points. Some papers which may help you are:
http://www.sciencedirect.com/science/article/pii/S0167865504003150
http://link.springer.com/chapter/10.1007/3-540-44692-3_30
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.59.8092&rep=rep1&type=pdf
http://proceedings.spiedigitallibrary.org/proceeding.aspx?articleid=906741
http://www.sciencedirect.com/science/article/pii/S0031320312004219
@John: I don't have any papers on that.
I attached the Delaunay triangulation of the white points in your segmentation. By varying the allowed maximum edge-length, you get one or several clusters. Allowing 'infinite' edge-lengths will get you the convex hull (the outer rim of attached image).
You could look at fuzzy connectedness ("Fuzzy Connectedness and Object Definition: Theory, Algorithms, and Applications in Image Segmentation", 1996) as a possible approach.
Another way would be to calculate the (Euclidean) Distance Transform: the value of each pixel is its distance to a white pixel. Thresholding the resulting float image at an 'appropriate' value will get you some clusters.
I altered code from Dave Watson:
http://www.cs.cmu.edu/afs/cs/project/cmcl/link.iwarp/member/jrs/fem/mesh/watson/nnsort.c
The computational geometry package of Matlab also has an implementation.
- Badges
- Science topic